| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | from collections.abc import Iterable, Iterator, Mapping, Sequence |
| 6 | from typing import Annotated, Any |
| 7 | |
| 8 | import numpy as np |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | import torch.nn.functional as F |
| 12 | from einops import rearrange |
| 13 | from transformers.image_processing_utils import BatchFeature |
| 14 | |
| 15 | from vllm.config import ModelConfig, VllmConfig |
| 16 | from vllm.config.multimodal import MultiModalDummyOptions |
| 17 | from vllm.distributed import parallel_state |
| 18 | from vllm.distributed import utils as dist_utils |
| 19 | from vllm.inputs import MultiModalDataDict |
| 20 | from vllm.model_executor.layers.attention import MMEncoderAttention |
| 21 | from vllm.model_executor.layers.linear import ( |
| 22 | ColumnParallelLinear, |
| 23 | QKVParallelLinear, |
| 24 | ReplicatedLinear, |
| 25 | RowParallelLinear, |
| 26 | ) |
| 27 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 28 | from vllm.model_executor.models.interfaces import ( |
| 29 | MultiModalEmbeddings, |
| 30 | SupportsLoRA, |
| 31 | SupportsMRoPE, |
| 32 | SupportsMultiModal, |
| 33 | SupportsPP, |
| 34 | ) |
| 35 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 36 | from vllm.model_executor.models.siglip import SiglipMLP |
| 37 | from vllm.model_executor.models.utils import ( |
| 38 | AutoWeightsLoader, |
| 39 | WeightsMapper, |
| 40 | init_vllm_registered_model, |
| 41 | maybe_prefix, |
| 42 | ) |
| 43 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 44 | from vllm.multimodal.inputs import ( |
| 45 | MultiModalFeatureSpec, |
| 46 | MultiModalFieldConfig, |
| 47 | MultiModalKwargsItems, |
| 48 | ) |
| 49 | from vllm.multimodal.parse import ImageSize, MultiModalDataItems |
| 50 | from vllm.multimodal.processing import ( |
| 51 | BaseDummyInputsBuilder, |
| 52 | BaseMultiModalProcessor, |
| 53 | BaseProcessingInfo, |
| 54 | PromptReplacement, |
| 55 | PromptUpdate, |
| 56 | PromptUpdateDetails, |
| 57 | cached_encode, |
| 58 | ) |
| 59 | from vllm.sequence import IntermediateTensors |
| 60 | from vllm.tokenizers import cached_tokenizer_from_config |
| 61 | from vllm.transformers_utils.config import patch_rope_parameters |
| 62 | from vllm.transformers_utils.configs.isaac import ( |
| 63 | IsaacConfig, |
| 64 | PixelShuffleSiglip2VisionConfig, |
| 65 | ) |
| 66 | from vllm.transformers_utils.processors.isaac import ( |
| 67 | IsaacImageProcessor, |
| 68 | IsaacProcessor, |
| 69 | get_image_size_for_max_num_patches, |
| 70 | ) |
| 71 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 72 | from vllm.utils.torch_utils import PIN_MEMORY, async_tensor_h2d |
| 73 | |
| 74 | from .vision import is_vit_use_data_parallel |
| 75 | |
| 76 | |
| 77 | def create_cumulative_seq_lengths( |
| 78 | seq_sizes: torch.Tensor, device: torch.device |
| 79 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 80 | """Create cumulative sequence lengths for variable-length attention.""" |
| 81 | cu_seqlens = torch.zeros(len(seq_sizes) + 1, dtype=torch.int32, device=device) |
| 82 | cu_seqlens[1:] = seq_sizes.cumsum(0) |
| 83 | max_seqlen = ( |
| 84 | seq_sizes.max() |
| 85 | if len(seq_sizes) > 0 |
| 86 | else torch.tensor(0, dtype=torch.int32, device=device) |
| 87 | ) |
| 88 | return cu_seqlens, max_seqlen |
| 89 | |
| 90 | |
| 91 | class Siglip2VariableSequenceEmbeddings(nn.Module): |
| 92 | def __init__(self, config: PixelShuffleSiglip2VisionConfig): |
| 93 | super().__init__() |
| 94 | self.config = config |
| 95 | self.embed_dim = config.hidden_size |
| 96 | self.patch_size = config.patch_size |
| 97 | |
| 98 | self.patch_embedding = ReplicatedLinear( |
| 99 | input_size=config.num_channels * self.patch_size * self.patch_size, |
| 100 | output_size=self.embed_dim, |
| 101 | return_bias=False, |
| 102 | ) |
| 103 | |
| 104 | self.num_patches = config.num_patches |
| 105 | self.position_embedding_size = int(self.num_patches**0.5) |
| 106 | self.position_embedding = nn.Embedding(self.num_patches, self.embed_dim) |
| 107 | |
| 108 | def positional_embeddings( |
| 109 | self, packed_seq_patches: tuple[torch.Tensor, torch.Tensor, torch.Tensor] |
| 110 | ) -> torch.Tensor: |
| 111 | # Prepare positional embeddings grid: (1, embed_dim, h, w) |
| 112 | positional_embeddings = ( |
| 113 | self.position_embedding.weight.reshape( |
| 114 | self.position_embedding_size, self.position_embedding_size, -1 |
| 115 | ) |
| 116 | .permute(2, 0, 1) |
| 117 | .unsqueeze(0) |
| 118 | ) |
| 119 | |
| 120 | _seq_patches, _seq_sizes, spatial_shapes = packed_seq_patches |
| 121 | pos_embeds_list = [] |
| 122 | mode = "bilinear" |
| 123 | align_corners = False |
| 124 | antialias = True |
| 125 | for spatial_shape in spatial_shapes: |
| 126 | height, width = int(spatial_shape[0]), int(spatial_shape[1]) |
| 127 | # Guard to ensure height and width are positive for torch.compile |
| 128 | if height > 0 and width > 0: |
| 129 | resized_pos_embed = F.interpolate( |
| 130 | positional_embeddings, |
| 131 | size=(height, width), |
| 132 | mode=mode, |
| 133 | align_corners=align_corners, |
| 134 | antialias=antialias, |
| 135 | ) |
| 136 | # Reshape from (1, embed_dim, height, width) to |
| 137 | # (height*width, embed_dim) |
| 138 | resized_pos_embed = resized_pos_embed.reshape( |
| 139 | self.embed_dim, height * width |
| 140 | ).transpose(0, 1) |
| 141 | else: |
| 142 | # Fallback - should never happen in practice |
| 143 | resized_pos_embed = positional_embeddings.reshape( |
| 144 | self.embed_dim, |
| 145 | self.position_embedding_size * self.position_embedding_size, |
| 146 | ).transpose(0, 1)[: height * width] |
| 147 | pos_embeds_list.append(resized_pos_embed) |
| 148 | |
| 149 | # Concatenate all positional embeddings along the sequence dimension |
| 150 | pos_embeds = torch.cat(pos_embeds_list, dim=0) |
| 151 | return pos_embeds |
| 152 | |
| 153 | def forward( |
| 154 | self, packed_seq_patches: tuple[torch.Tensor, torch.Tensor, torch.Tensor] |
| 155 | ): |
| 156 | seq_patches, _seq_sizes, _spatial_shapes = packed_seq_patches |
| 157 | |
| 158 | target_weight = self.patch_embedding.weight |
| 159 | seq_patches = seq_patches.to( |
| 160 | device=target_weight.device, dtype=target_weight.dtype |
| 161 | ) |
| 162 | patch_embeds = self.patch_embedding(seq_patches) |
| 163 | pos_embeds = self.positional_embeddings(packed_seq_patches) |
| 164 | |
| 165 | # Flatten patch embeddings to match positional embeddings format |
| 166 | if patch_embeds.dim() == 3: |
| 167 | patch_embeds = patch_embeds.view(-1, patch_embeds.size(-1)) |
| 168 | |
| 169 | # Add positional embeddings to patch embeddings |
| 170 | embeddings = patch_embeds + pos_embeds |
| 171 | return embeddings |
| 172 | |
| 173 | |
| 174 | def create_pixel_shuffle_index_map( |
| 175 | seq_sizes: torch.Tensor, |
| 176 | token_grids: torch.Tensor, |
| 177 | scale_factor: int = 1, |
| 178 | device: torch.device | None = None, |
| 179 | ) -> torch.Tensor: |
| 180 | """Build a gather-index map that tells us, for every *output* token after |
| 181 | pixel-shuffle, which `scale_factor**2` *input* tokens are being merged. |
| 182 | |
| 183 | Args: |
| 184 | ---- |
| 185 | seq_sizes : (num_images,) - #patches in each image (row-major order) |
| 186 | token_grids : (num_images,2) - (height, width) for every image |
| 187 | scale_factor : spatial down-scale factor (≥2) |
| 188 | device : (optional) overrides `seq_sizes.device` |
| 189 | |
| 190 | Returns: |
| 191 | ------- |
| 192 | gather_idx : (new_total_seq_len, scale_factor**2) int64 tensor. |
| 193 | gather_idx[i, j] is the *flat* index into the *original* |
| 194 | packed sequence for the j-th sub-patch that forms the |
| 195 | i-th output token. |
| 196 | |
| 197 | """ |
| 198 | if device is None: |
| 199 | device = seq_sizes.device |
| 200 | |
| 201 | r = int(scale_factor) |
| 202 | if r < 2: |
| 203 | raise ValueError("`scale_factor` must be ≥ 2") |
| 204 | |
| 205 | # Safety: all spatial dims must be divisible by r |
| 206 | # Cannot run under torch compile fullgraph mode hence |
| 207 | if not torch.compiler.is_compiling() and not ( |
| 208 | (token_grids[:, 0] % r == 0).all() and (token_grids[:, 1] % r == 0).all() |
| 209 | ): |
| 210 | raise AssertionError( |
| 211 | "Every (H,W) in `token_grids` must be divisible by " |
| 212 | f"scale_factor={r}, got {token_grids.tolist()}" |
| 213 | ) |
| 214 | |
| 215 | gather_chunks: list[torch.Tensor] = [] |
| 216 | tok_offset = 0 |
| 217 | |
| 218 | for seq_len, (h, w) in zip(seq_sizes.tolist(), token_grids.tolist(), strict=False): |
| 219 | # Build the (H, W) grid of flat indices for this image |
| 220 | grid = torch.arange(seq_len, device=device, dtype=torch.int64) + tok_offset |
| 221 | grid = grid.view(h, w) # (H, W) |
| 222 | |
| 223 | # -------- identical ordering to your fixed-res routine -------- |
| 224 | # Step 1: split width into blocks of r |
| 225 | grid = grid.view(h, w // r, r) # (H, W/r, r) |
| 226 | # Step 2: now split height into blocks of r |
| 227 | grid = grid.view(h // r, r, w // r, r) # (H/r, r, W/r, r) |
| 228 | # Step 3: final permutation to (H/r, W/r, r, r) |
| 229 | grid = grid.permute(0, 2, 1, 3).contiguous() # (H/r, W/r, r, r) |
| 230 | # Step 4: each (r, r) block forms one output token |
| 231 | gather_chunks.append(grid.reshape(-1, r * r)) # (H*W / r², r²) |
| 232 | |
| 233 | tok_offset += seq_len |
| 234 | |
| 235 | # Concatenate over all images in the packed batch |
| 236 | gather_idx = torch.cat(gather_chunks, dim=0) # (Σ_i HᵢWᵢ/r², r²) |
| 237 | return gather_idx |
| 238 | |
| 239 | |
| 240 | def pixel_shuffle_varlen( |
| 241 | x: torch.Tensor, |
| 242 | token_grids: torch.Tensor, |
| 243 | scale_factor: int = 1, |
| 244 | ) -> torch.Tensor: |
| 245 | r"""Apply pixel shuffle to a packed vision sequence without unpacking per image. |
| 246 | |
| 247 | Args: |
| 248 | x (`torch.Tensor`): |
| 249 | Concatenated vision embeddings. Accepts `(seq_len, hidden_size)` or |
| 250 | `(1, seq_len, hidden_size)` shapes produced by stacking image |
| 251 | patches. |
| 252 | token_grids (`torch.Tensor`): |
| 253 | Integer tensor of shape `(num_images, 2)` whose rows give the |
| 254 | `(height, width)` patch grid sizes corresponding to each image |
| 255 | segment inside `x`. |
| 256 | scale_factor (`int`, *optional*, defaults to 1): |
| 257 | Spatial down-sampling factor specific to pixel shuffle. Values |
| 258 | greater than one merge `scale_factor**2` neighboring patches into a |
| 259 | single embedding channel-group. |
| 260 | |
| 261 | Returns: |
| 262 | `torch.Tensor`: Pixel-shuffled embeddings with shape matching the input |
| 263 | convention: `(seq_len, hidden_size * scale_factor**2)` when the input |
| 264 | was 2D, or `(1, seq_len, hidden_size * scale_factor**2)` if the |
| 265 | singleton batch dimension was present. |
| 266 | |
| 267 | Raises: |
| 268 | ValueError: If more than one batch item is provided. |
| 269 | |
| 270 | """ |
| 271 | keep_batch_dim = x.dim() == 3 |
| 272 | if keep_batch_dim: |
| 273 | if x.size(0) != 1: |
| 274 | raise AssertionError("Packed sequence is expected to have batch_size == 1") |
| 275 | x_ = x.squeeze(0) # (seq, embed) |
| 276 | else: |
| 277 | x_ = x # (seq, embed) |
| 278 | |
| 279 | embed_dim = x_.size(-1) |
| 280 | r = int(scale_factor) |
| 281 | |
| 282 | # Calculate seq_sizes from token_grids |
| 283 | seq_sizes = torch.prod(token_grids, dim=-1) |
| 284 | |
| 285 | # Build index map and gather in one go |
| 286 | gather_idx = create_pixel_shuffle_index_map( |
| 287 | seq_sizes=seq_sizes, |
| 288 | token_grids=token_grids, |
| 289 | scale_factor=r, |
| 290 | device=x_.device, |
| 291 | ) # (new_seq, r²) |
| 292 | |
| 293 | # Gather → (new_seq, r², embed_dim) |
| 294 | gathered = x_[gather_idx] # fancy indexing keeps gradient |
| 295 | |
| 296 | # Merge the r² group dimension into channels to finish the shuffle |
| 297 | out = gathered.reshape(gathered.size(0), embed_dim * r * r) |
| 298 | |
| 299 | # Restore batch dimension if needed |
| 300 | if keep_batch_dim: |
| 301 | out = out.unsqueeze(0) |
| 302 | return out |
| 303 | |
| 304 | |
| 305 | # ============================================================================ |
| 306 | # Configuration |
| 307 | # ============================================================================ |
| 308 | |
| 309 | |
| 310 | class IsaacProcessingInfo(BaseProcessingInfo): |
| 311 | def get_hf_config(self) -> IsaacConfig: |
| 312 | if hasattr(self.ctx, "get_hf_config"): |
| 313 | original_config = self.ctx.get_hf_config() |
| 314 | # Map HF config parameters to our vLLM config parameters |
| 315 | return IsaacConfig( |
| 316 | # Vision parameters - map from HF names |
| 317 | vision_config=getattr(original_config, "vision_config", None), |
| 318 | vision_patch_size=getattr(original_config, "video_patch_size", 16), |
| 319 | vision_max_num_patches=getattr( |
| 320 | original_config, "vision_max_num_patches", 256 |
| 321 | ), |
| 322 | vision_min_num_patches=getattr( |
| 323 | original_config, "vision_min_num_patches", None |
| 324 | ), |
| 325 | pixel_shuffle_scale=getattr(original_config, "pixel_shuffle_scale", 1), |
| 326 | max_sequence_length=getattr( |
| 327 | original_config, "max_sequence_length", 16384 |
| 328 | ), |
| 329 | vision_token=getattr(original_config, "vision_token", "<image>"), |
| 330 | vision_attn_implementation=getattr( |
| 331 | original_config, "vision_attn_implementation", None |
| 332 | ), |
| 333 | ) |
| 334 | return IsaacConfig() |
| 335 | |
| 336 | def get_image_processor(self, **kwargs) -> IsaacImageProcessor: |
| 337 | return IsaacImageProcessor(**kwargs) |
| 338 | |
| 339 | def get_hf_processor(self, **kwargs) -> IsaacProcessor: |
| 340 | hf_config = self.get_hf_config() |
| 341 | |
| 342 | return IsaacProcessor( |
| 343 | tokenizer=self.get_tokenizer(), |
| 344 | image_processor=self.get_image_processor(**kwargs), |
| 345 | image_token=hf_config.vision_token, |
| 346 | ) |
| 347 | |
| 348 | def get_image_size_with_most_features(self) -> ImageSize: |
| 349 | hf_config = self.get_hf_config() |
| 350 | # Get target dimensions |
| 351 | target_height, target_width = get_image_size_for_max_num_patches( |
| 352 | 9999999, |
| 353 | 9999999, |
| 354 | hf_config.video_patch_size, |
| 355 | hf_config.vision_max_num_patches, |
| 356 | min_num_patches=hf_config.vision_min_num_patches, |
| 357 | pixel_shuffle_scale=hf_config.pixel_shuffle_scale, |
| 358 | ) |
| 359 | return ImageSize(width=target_width, height=target_height) |
| 360 | |
| 361 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 362 | return {"image": None} |
| 363 | |
| 364 | def get_mm_max_tokens_per_item( |
| 365 | self, |
| 366 | seq_len: int, |
| 367 | mm_counts: Mapping[str, int], |
| 368 | ) -> Mapping[str, int]: |
| 369 | hf_config = self.get_hf_config() |
| 370 | num_vision_tokens = hf_config.vision_max_num_patches // ( |
| 371 | hf_config.pixel_shuffle_scale**2 |
| 372 | ) |
| 373 | return {"image": num_vision_tokens} |
| 374 | |
| 375 | |
| 376 | class IsaacDummyInputsBuilder(BaseDummyInputsBuilder[IsaacProcessingInfo]): |
| 377 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 378 | num_images = mm_counts.get("image", 0) |
| 379 | |
| 380 | hf_processor = self.info.get_hf_processor() |
| 381 | image_token: str = hf_processor.image_token |
| 382 | |
| 383 | return image_token * num_images |
| 384 | |
| 385 | def get_dummy_mm_data( |
| 386 | self, |
| 387 | seq_len: int, |
| 388 | mm_counts: Mapping[str, int], |
| 389 | mm_options: MultiModalDummyOptions, |
| 390 | ) -> MultiModalDataDict: |
| 391 | target_width, target_height = self.info.get_image_size_with_most_features() |
| 392 | |
| 393 | return { |
| 394 | "image": self._get_dummy_images( |
| 395 | width=target_width, |
| 396 | height=target_height, |
| 397 | num_images=mm_counts.get("image", 0), |
| 398 | overrides=mm_options.get("image"), |
| 399 | ), |
| 400 | } |
| 401 | |
| 402 | |
| 403 | class IsaacImagePixelInputs(TensorSchema): |
| 404 | """Schema for validating Isaac image inputs. |
| 405 | |
| 406 | Dimensions: |
| 407 | - np: Number of patches |
| 408 | - d: Patch dimension |
| 409 | - ni: Number of images |
| 410 | |
| 411 | The schema enforces: |
| 412 | - pixel_values must be 2D: (num_patches, patch_dim) |
| 413 | - image_grid_thw must be 2D: (num_images, 3) |
| 414 | where 3 represents [T, H, W] |
| 415 | """ |
| 416 | |
| 417 | pixel_values: Annotated[ |
| 418 | torch.Tensor, |
| 419 | TensorShape("np", "d"), |
| 420 | ] |
| 421 | |
| 422 | image_grid_thw: Annotated[ |
| 423 | torch.Tensor, |
| 424 | TensorShape("ni", 3), |
| 425 | ] |
| 426 | |
| 427 | |
| 428 | class IsaacMultiModalProcessor(BaseMultiModalProcessor): |
| 429 | def _get_mm_fields_config( |
| 430 | self, |
| 431 | hf_inputs: BatchFeature, |
| 432 | hf_processor_mm_kwargs: Mapping[str, object], |
| 433 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 434 | # Configure multimodal fields for Isaac model |
| 435 | image_grid_thw = hf_inputs.get("image_grid_thw", torch.empty((0, 3))) |
| 436 | image_grid_sizes = image_grid_thw.prod(-1) |
| 437 | |
| 438 | return { |
| 439 | "pixel_values": MultiModalFieldConfig.flat_from_sizes( |
| 440 | "image", image_grid_sizes |
| 441 | ), |
| 442 | "image_grid_thw": MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 443 | } |
| 444 | |
| 445 | def _get_prompt_updates( |
| 446 | self, |
| 447 | mm_items: MultiModalDataItems, |
| 448 | hf_processor_mm_kwargs: Mapping[str, Any], |
| 449 | out_mm_kwargs: MultiModalKwargsItems, |
| 450 | ) -> Sequence[PromptUpdate]: |
| 451 | image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs) |
| 452 | tokenizer = self.info.get_tokenizer() |
| 453 | image_pad_token_ids = cached_encode( |
| 454 | tokenizer, "<|image_pad|>", add_special_tokens=False |
| 455 | ) |
| 456 | |
| 457 | pixel_shuffle_scale = getattr(image_processor, "pixel_shuffle_scale", 2) |
| 458 | merge_length = pixel_shuffle_scale**2 |
| 459 | |
| 460 | def get_replacement_isaac(item_idx: int): |
| 461 | out_item = out_mm_kwargs["image"][item_idx] |
| 462 | grid_thw = out_item["image_grid_thw"].data |
| 463 | assert isinstance(grid_thw, torch.Tensor) |
| 464 | |
| 465 | feature_size = int(grid_thw.prod()) // merge_length |
| 466 | repl_full = image_pad_token_ids * feature_size |
| 467 | return PromptUpdateDetails.select_token_ids(repl_full, image_pad_token_ids) |
| 468 | |
| 469 | return [ |
| 470 | PromptReplacement( |
| 471 | modality="image", |
| 472 | target=cached_encode(tokenizer, "<image>", add_special_tokens=False), |
| 473 | replacement=get_replacement_isaac, |
| 474 | ) |
| 475 | ] |
| 476 | |
| 477 | |
| 478 | class Siglip2VisionAttention(nn.Module): |
| 479 | def __init__( |
| 480 | self, |
| 481 | config: PixelShuffleSiglip2VisionConfig, |
| 482 | quant_config: QuantizationConfig | None = None, |
| 483 | *, |
| 484 | prefix: str = "", |
| 485 | ) -> None: |
| 486 | super().__init__() |
| 487 | |
| 488 | use_data_parallel = is_vit_use_data_parallel() |
| 489 | self.tp_size = ( |
| 490 | 1 |
| 491 | if use_data_parallel |
| 492 | else parallel_state.get_tensor_model_parallel_world_size() |
| 493 | ) |
| 494 | self.tp_rank = parallel_state.get_tensor_model_parallel_rank() |
| 495 | self.hidden_size_per_attention_head = dist_utils.divide( |
| 496 | config.hidden_size, config.num_attention_heads |
| 497 | ) |
| 498 | self.num_attention_heads_per_partition = dist_utils.divide( |
| 499 | config.num_attention_heads, self.tp_size |
| 500 | ) |
| 501 | |
| 502 | self.qkv_proj = QKVParallelLinear( |
| 503 | hidden_size=config.hidden_size, |
| 504 | head_size=self.hidden_size_per_attention_head, |
| 505 | total_num_heads=config.num_attention_heads, |
| 506 | total_num_kv_heads=config.num_attention_heads, |
| 507 | bias=True, |
| 508 | quant_config=quant_config, |
| 509 | prefix=f"{prefix}.qkv_proj", |
| 510 | disable_tp=use_data_parallel, |
| 511 | ) |
| 512 | self.out_proj = RowParallelLinear( |
| 513 | input_size=config.hidden_size, |
| 514 | output_size=config.hidden_size, |
| 515 | quant_config=quant_config, |
| 516 | prefix=f"{prefix}.out_proj", |
| 517 | disable_tp=use_data_parallel, |
| 518 | ) |
| 519 | |
| 520 | self.attn = MMEncoderAttention( |
| 521 | num_heads=self.num_attention_heads_per_partition, |
| 522 | head_size=self.hidden_size_per_attention_head, |
| 523 | scale=self.hidden_size_per_attention_head**-0.5, |
| 524 | prefix=f"{prefix}.attn", |
| 525 | ) |
| 526 | |
| 527 | def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]: |
| 528 | seq_len, bs, _ = qkv.shape |
| 529 | q, k, v = qkv.chunk(3, dim=2) |
| 530 | new_shape = ( |
| 531 | seq_len, |
| 532 | bs, |
| 533 | self.num_attention_heads_per_partition, |
| 534 | self.hidden_size_per_attention_head, |
| 535 | ) |
| 536 | q, k, v = (x.view(*new_shape) for x in (q, k, v)) |
| 537 | return q, k, v |
| 538 | |
| 539 | def forward( |
| 540 | self, |
| 541 | hidden_states: torch.Tensor, |
| 542 | *, |
| 543 | cu_seqlens: torch.Tensor, |
| 544 | max_seqlen: torch.Tensor | None, |
| 545 | ) -> torch.Tensor: |
| 546 | batch_size, _, _ = hidden_states.shape |
| 547 | if batch_size != 1: |
| 548 | raise ValueError("packed variable-length attention expects batch_size=1") |
| 549 | x = rearrange(hidden_states, "b s d -> s b d") |
| 550 | x, _ = self.qkv_proj(x) |
| 551 | q, k, v = self.split_qkv(x) |
| 552 | q, k, v = (rearrange(t, "s b h d -> b s h d") for t in (q, k, v)) |
| 553 | |
| 554 | context_layer = self.attn( |
| 555 | query=q, |
| 556 | key=k, |
| 557 | value=v, |
| 558 | cu_seqlens=cu_seqlens, |
| 559 | max_seqlen=max_seqlen, |
| 560 | ) |
| 561 | context_layer = rearrange(context_layer, "b s h d -> s b (h d)").contiguous() |
| 562 | |
| 563 | output, _ = self.out_proj(context_layer) |
| 564 | output = rearrange(output, "s b d -> b s d") |
| 565 | return output |
| 566 | |
| 567 | |
| 568 | class Siglip2EncoderLayer(nn.Module): |
| 569 | def __init__( |
| 570 | self, |
| 571 | config: PixelShuffleSiglip2VisionConfig, |
| 572 | quant_config: QuantizationConfig | None = None, |
| 573 | *, |
| 574 | prefix: str = "", |
| 575 | ) -> None: |
| 576 | super().__init__() |
| 577 | self.embed_dim = config.hidden_size |
| 578 | self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) |
| 579 | self.self_attn = Siglip2VisionAttention( |
| 580 | config, |
| 581 | quant_config=quant_config, |
| 582 | prefix=f"{prefix}.self_attn", |
| 583 | ) |
| 584 | self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) |
| 585 | self.mlp = SiglipMLP( |
| 586 | config, |
| 587 | quant_config=quant_config, |
| 588 | prefix=f"{prefix}.mlp", |
| 589 | ) |
| 590 | |
| 591 | def forward( |
| 592 | self, |
| 593 | hidden_states: torch.Tensor, |
| 594 | *, |
| 595 | cu_seqlens: torch.Tensor, |
| 596 | max_seqlen: torch.Tensor | None, |
| 597 | ) -> torch.Tensor: |
| 598 | residual = hidden_states |
| 599 | |
| 600 | hidden_states = self.layer_norm1(hidden_states) |
| 601 | hidden_states = self.self_attn( |
| 602 | hidden_states=hidden_states, |
| 603 | cu_seqlens=cu_seqlens, |
| 604 | max_seqlen=max_seqlen, |
| 605 | ) |
| 606 | hidden_states = residual + hidden_states |
| 607 | |
| 608 | residual = hidden_states |
| 609 | hidden_states = self.layer_norm2(hidden_states) |
| 610 | hidden_states = self.mlp(hidden_states) |
| 611 | hidden_states = residual + hidden_states |
| 612 | |
| 613 | return hidden_states |
| 614 | |
| 615 | |
| 616 | class Siglip2Encoder(nn.Module): |
| 617 | def __init__( |
| 618 | self, |
| 619 | config: PixelShuffleSiglip2VisionConfig, |
| 620 | quant_config: QuantizationConfig | None = None, |
| 621 | *, |
| 622 | prefix: str = "", |
| 623 | ) -> None: |
| 624 | super().__init__() |
| 625 | self.config = config |
| 626 | self.layers = nn.ModuleList( |
| 627 | [ |
| 628 | Siglip2EncoderLayer( |
| 629 | config, |
| 630 | quant_config=quant_config, |
| 631 | prefix=f"{prefix}.layers.{layer_idx}", |
| 632 | ) |
| 633 | for layer_idx in range(config.num_hidden_layers) |
| 634 | ] |
| 635 | ) |
| 636 | |
| 637 | def forward( |
| 638 | self, |
| 639 | inputs_embeds: torch.Tensor, |
| 640 | *, |
| 641 | cu_seqlens: torch.Tensor | None = None, |
| 642 | max_seqlen: torch.Tensor | None = None, |
| 643 | ) -> torch.Tensor: |
| 644 | hidden_states = inputs_embeds |
| 645 | for encoder_layer in self.layers: |
| 646 | hidden_states = encoder_layer( |
| 647 | hidden_states, |
| 648 | cu_seqlens=cu_seqlens, |
| 649 | max_seqlen=max_seqlen, |
| 650 | ) |
| 651 | return hidden_states |
| 652 | |
| 653 | |
| 654 | class Siglip2VisionTransformer(nn.Module): |
| 655 | hf_to_vllm_mapper = WeightsMapper( |
| 656 | orig_to_new_stacked={ |
| 657 | ".q_proj": (".qkv_proj", "q"), |
| 658 | ".k_proj": (".qkv_proj", "k"), |
| 659 | ".v_proj": (".qkv_proj", "v"), |
| 660 | } |
| 661 | ) |
| 662 | |
| 663 | def __init__( |
| 664 | self, |
| 665 | config: PixelShuffleSiglip2VisionConfig, |
| 666 | quant_config: QuantizationConfig | None = None, |
| 667 | prefix: str = "", |
| 668 | ): |
| 669 | super().__init__() |
| 670 | self.config = config |
| 671 | self.quant_config = quant_config |
| 672 | embed_dim = config.hidden_size |
| 673 | |
| 674 | self.embeddings = Siglip2VariableSequenceEmbeddings(config) |
| 675 | self.pixel_shuffle_scale_factor = config.pixel_shuffle_scale_factor |
| 676 | self.encoder = Siglip2Encoder( |
| 677 | config, |
| 678 | quant_config=quant_config, |
| 679 | prefix=f"{prefix}.encoder", |
| 680 | ) |
| 681 | self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) |
| 682 | |
| 683 | def forward( |
| 684 | self, |
| 685 | packed_seq_patches: tuple[torch.Tensor, torch.Tensor], |
| 686 | ) -> torch.Tensor: |
| 687 | r"""spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`): |
| 688 | Tensor containing the spatial dimensions (height, width) |
| 689 | of the input images. |
| 690 | """ |
| 691 | seq_patches, token_grids = packed_seq_patches |
| 692 | seq_sizes = torch.prod(token_grids, dim=-1) |
| 693 | |
| 694 | # Get embeddings from packed sequence |
| 695 | hidden_states = self.embeddings((seq_patches, seq_sizes, token_grids)) |
| 696 | |
| 697 | # Add a pseudo batch dimension for the encoder |
| 698 | hidden_states = hidden_states.unsqueeze(0) |
| 699 | |
| 700 | cu_seqlens, max_seqlen = create_cumulative_seq_lengths( |
| 701 | seq_sizes, hidden_states.device |
| 702 | ) |
| 703 | |
| 704 | hidden_states = self.encoder( |
| 705 | inputs_embeds=hidden_states, |
| 706 | cu_seqlens=cu_seqlens, |
| 707 | max_seqlen=max_seqlen, |
| 708 | ) |
| 709 | hidden_states = self.post_layernorm(hidden_states) |
| 710 | |
| 711 | if self.pixel_shuffle_scale_factor > 1: |
| 712 | hidden_states = pixel_shuffle_varlen( |
| 713 | x=hidden_states, |
| 714 | token_grids=token_grids, |
| 715 | scale_factor=self.pixel_shuffle_scale_factor, |
| 716 | ) |
| 717 | # Remove the pseudo batch dimension we added earlier |
| 718 | hidden_states = hidden_states.squeeze(0) |
| 719 | |
| 720 | # return last_hidden_state |
| 721 | return hidden_states |
| 722 | |
| 723 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 724 | loader = AutoWeightsLoader(self) |
| 725 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 726 | |
| 727 | |
| 728 | def _resolve_vision_token_id(model_config: ModelConfig, vision_token: str) -> int: |
| 729 | tokenizer = cached_tokenizer_from_config(model_config) |
| 730 | assert tokenizer is not None |
| 731 | return tokenizer.encode(vision_token, add_special_tokens=False)[0] |
| 732 | |
| 733 | |
| 734 | class IsaacVisionEmbedding(nn.Module): |
| 735 | def __init__( |
| 736 | self, |
| 737 | vision_cfg: PixelShuffleSiglip2VisionConfig, |
| 738 | hidden_dim: int, |
| 739 | output_dim: int, |
| 740 | quant_config: QuantizationConfig | None = None, |
| 741 | prefix: str = "", |
| 742 | ): |
| 743 | super().__init__() |
| 744 | self.transformer = Siglip2VisionTransformer( |
| 745 | vision_cfg, |
| 746 | quant_config=quant_config, |
| 747 | prefix=maybe_prefix(prefix, "0"), |
| 748 | ) |
| 749 | self.linear_fc1 = ColumnParallelLinear( |
| 750 | hidden_dim, |
| 751 | 4 * hidden_dim, |
| 752 | bias=False, |
| 753 | quant_config=quant_config, |
| 754 | prefix=maybe_prefix(prefix, "1"), |
| 755 | return_bias=False, |
| 756 | ) |
| 757 | self.act = nn.SiLU() |
| 758 | self.linear_fc2 = RowParallelLinear( |
| 759 | 4 * hidden_dim, |
| 760 | output_dim, |
| 761 | bias=False, |
| 762 | quant_config=quant_config, |
| 763 | prefix=maybe_prefix(prefix, "3"), |
| 764 | return_bias=False, |
| 765 | ) |
| 766 | |
| 767 | def forward( |
| 768 | self, packed_seq_patches: tuple[torch.Tensor, torch.Tensor] |
| 769 | ) -> torch.Tensor: |
| 770 | hidden_states = self.transformer(packed_seq_patches) |
| 771 | hidden_states = self.linear_fc1(hidden_states) |
| 772 | hidden_states = self.act(hidden_states) |
| 773 | hidden_states = self.linear_fc2(hidden_states) |
| 774 | return hidden_states |
| 775 | |
| 776 | |
| 777 | @MULTIMODAL_REGISTRY.register_processor( |
| 778 | IsaacMultiModalProcessor, |
| 779 | info=IsaacProcessingInfo, |
| 780 | dummy_inputs=IsaacDummyInputsBuilder, |
| 781 | ) |
| 782 | class IsaacForConditionalGeneration( |
| 783 | nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE |
| 784 | ): |
| 785 | packed_modules_mapping = { |
| 786 | "qkv_proj": [ |
| 787 | "q_proj", |
| 788 | "k_proj", |
| 789 | "v_proj", |
| 790 | ], |
| 791 | "gate_up_proj": [ |
| 792 | "gate_proj", |
| 793 | "up_proj", |
| 794 | ], |
| 795 | } |
| 796 | |
| 797 | supports_encoder_tp_data = True |
| 798 | |
| 799 | # To ensure correct weight loading and mapping. |
| 800 | hf_to_vllm_mapper = WeightsMapper( |
| 801 | orig_to_new_prefix={ |
| 802 | "lm_head.": "language_model.lm_head.", |
| 803 | "model.text_model.lm_head.": "language_model.lm_head.", |
| 804 | "model.text_model.": "language_model.model.", |
| 805 | "model.vision_embedding.0": "vision_embedding.transformer", |
| 806 | "model.vision_embedding.1": "vision_embedding.linear_fc1", |
| 807 | "model.vision_embedding.2": "vision_embedding.act", |
| 808 | "model.vision_embedding.3": "vision_embedding.linear_fc2", |
| 809 | "model.vision_embedding.": "vision_embedding.", |
| 810 | "model.lm_head.": "language_model.lm_head.", |
| 811 | "model.": "language_model.model.", |
| 812 | } |
| 813 | ) |
| 814 | |
| 815 | @classmethod |
| 816 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 817 | if modality.startswith("image"): |
| 818 | return "<image>" |
| 819 | |
| 820 | raise ValueError("Only image modality is supported") |
| 821 | |
| 822 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "model"): |
| 823 | super().__init__() |
| 824 | config: IsaacConfig = vllm_config.model_config.hf_config |
| 825 | quant_config = vllm_config.quant_config |
| 826 | self.config = config |
| 827 | |
| 828 | head_dim = config.head_dim |
| 829 | calculated_mrope_section = [ |
| 830 | head_dim // 4, # 2x more for temporal dim |
| 831 | head_dim // 8, |
| 832 | head_dim // 8, |
| 833 | ] |
| 834 | |
| 835 | self.vision_token_id = _resolve_vision_token_id( |
| 836 | vllm_config.model_config, config.vision_token |
| 837 | ) |
| 838 | config.image_token_id = self.vision_token_id |
| 839 | |
| 840 | text_cfg = getattr(config, "text_config", None) |
| 841 | target_cfg = ( |
| 842 | text_cfg |
| 843 | if text_cfg is not None and not isinstance(text_cfg, dict) |
| 844 | else config |
| 845 | ) |
| 846 | |
| 847 | rope_scaling = getattr(target_cfg, "rope_scaling", None) |
| 848 | if rope_scaling is None and target_cfg is config: |
| 849 | rope_scaling = getattr(config, "_rope_scaling", None) |
| 850 | |
| 851 | patch_rope_parameters(target_cfg) |
| 852 | rope_parameters = target_cfg.rope_parameters |
| 853 | rope_parameters["mrope_section"] = calculated_mrope_section |
| 854 | if rope_scaling is not None and "mrope_interleaved" in rope_scaling: |
| 855 | rope_parameters.setdefault( |
| 856 | "mrope_interleaved", rope_scaling["mrope_interleaved"] |
| 857 | ) |
| 858 | target_cfg.rope_parameters = rope_parameters |
| 859 | |
| 860 | with self._mark_language_model(vllm_config): |
| 861 | self.language_model = init_vllm_registered_model( |
| 862 | vllm_config=vllm_config, |
| 863 | architectures=["Qwen3ForCausalLM"], |
| 864 | prefix=maybe_prefix(prefix, "language_model"), |
| 865 | ) |
| 866 | |
| 867 | self.make_empty_intermediate_tensors = ( |
| 868 | self.language_model.make_empty_intermediate_tensors |
| 869 | ) |
| 870 | |
| 871 | vision_cfg = config.vision_config |
| 872 | if vision_cfg is None: |
| 873 | raise ValueError("IsaacConfig should always have vision_config") |
| 874 | attn_impl = ( |
| 875 | config.vision_attn_implementation |
| 876 | if config.vision_attn_implementation is not None |
| 877 | else getattr(config, "_attn_implementation", None) |
| 878 | ) |
| 879 | if attn_impl is not None: |
| 880 | vision_cfg._attn_implementation = attn_impl |
| 881 | |
| 882 | hidden_dim = vision_cfg.hidden_size * (vision_cfg.pixel_shuffle_scale_factor**2) |
| 883 | |
| 884 | with self._mark_tower_model(vllm_config, "image"): |
| 885 | self.vision_embedding = IsaacVisionEmbedding( |
| 886 | vision_cfg=vision_cfg, |
| 887 | hidden_dim=hidden_dim, |
| 888 | output_dim=config.hidden_size, |
| 889 | quant_config=quant_config, |
| 890 | prefix=maybe_prefix(prefix, "vision_embedding"), |
| 891 | ) |
| 892 | |
| 893 | def iter_mm_grid_hw( |
| 894 | self, input_tokens: list[int], mm_features: list[MultiModalFeatureSpec] |
| 895 | ) -> Iterator[tuple[int, int, int]]: |
| 896 | spatial_merge_size = self.config.vision_config.pixel_shuffle_scale_factor |
| 897 | for mm_feature in sorted(mm_features, key=lambda f: f.mm_position.offset): |
| 898 | offset = mm_feature.mm_position.offset |
| 899 | if mm_feature.modality == "image": |
| 900 | mm_data = mm_feature.data |
| 901 | assert mm_data is not None |
| 902 | grid_thw = mm_data["image_grid_thw"].data |
| 903 | assert isinstance(grid_thw, torch.Tensor) |
| 904 | t, h, w = grid_thw.tolist() |
| 905 | assert t == 1, f"Image must have 1 frame, got {t}" |
| 906 | yield offset, h // spatial_merge_size, w // spatial_merge_size |
| 907 | else: |
| 908 | raise ValueError(f"Unsupported modality: {mm_feature.modality}") |
| 909 | |
| 910 | def get_mrope_input_positions( |
| 911 | self, |
| 912 | input_tokens: list[int], |
| 913 | mm_features: list[MultiModalFeatureSpec], |
| 914 | ) -> tuple[torch.Tensor, int]: |
| 915 | llm_pos_ids_list: list[np.ndarray] = [] |
| 916 | st = 0 |
| 917 | for offset, llm_grid_h, llm_grid_w in self.iter_mm_grid_hw( |
| 918 | input_tokens, mm_features |
| 919 | ): |
| 920 | text_len = offset - st |
| 921 | st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 |
| 922 | llm_pos_ids_list.append( |
| 923 | np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx |
| 924 | ) |
| 925 | |
| 926 | grid_indices = np.indices((1, llm_grid_h, llm_grid_w)).reshape(3, -1) |
| 927 | grid_indices[0, :] = grid_indices[0, :] + text_len + st_idx |
| 928 | llm_pos_ids_list.append(grid_indices) |
| 929 | st = offset + llm_grid_h * llm_grid_w |
| 930 | |
| 931 | if st < len(input_tokens): |
| 932 | st_idx = llm_pos_ids_list[-1][0, -1] + 1 if len(llm_pos_ids_list) > 0 else 0 |
| 933 | text_len = len(input_tokens) - st |
| 934 | llm_pos_ids_list.append( |
| 935 | np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx |
| 936 | ) |
| 937 | |
| 938 | llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1) |
| 939 | mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item() |
| 940 | |
| 941 | return torch.from_numpy(llm_positions), mrope_position_delta |
| 942 | |
| 943 | def _parse_and_validate_image_input( |
| 944 | self, **kwargs: object |
| 945 | ) -> IsaacImagePixelInputs | None: |
| 946 | pixel_values = kwargs.get("pixel_values") |
| 947 | image_grid_thw = kwargs.get("image_grid_thw") |
| 948 | if pixel_values is None or image_grid_thw is None: |
| 949 | return None |
| 950 | |
| 951 | # TensorSchema will automatically validate shapes on initialization |
| 952 | return IsaacImagePixelInputs( |
| 953 | pixel_values=pixel_values, |
| 954 | image_grid_thw=image_grid_thw, |
| 955 | ) |
| 956 | |
| 957 | def _process_image_input( |
| 958 | self, |
| 959 | image_input: IsaacImagePixelInputs, |
| 960 | ) -> tuple[torch.Tensor, ...]: |
| 961 | pixel_values = image_input["pixel_values"] |
| 962 | image_grid_thw = image_input["image_grid_thw"] |
| 963 | if pixel_values.numel() == 0: |
| 964 | return () |
| 965 | |
| 966 | device = next(self.language_model.parameters()).device |
| 967 | dtype = self.vision_embedding.linear_fc1.weight.dtype |
| 968 | pixel_values = pixel_values.to(device=device, dtype=dtype) |
| 969 | # The [:, 1:3] column slice isn't densely laid out, so stage it with a |
| 970 | # single copy straight into pinned memory to keep the H2D non-blocking. |
| 971 | spatial_grids_cpu = torch.empty( |
| 972 | (image_grid_thw.shape[0], 2), |
| 973 | dtype=image_grid_thw.dtype, |
| 974 | pin_memory=PIN_MEMORY, |
| 975 | ).copy_(image_grid_thw[:, 1:3]) |
| 976 | spatial_grids = async_tensor_h2d(spatial_grids_cpu, device, dtype=torch.int32) |
| 977 | |
| 978 | vision_embeddings = self.vision_embedding((pixel_values, spatial_grids)) |
| 979 | merge_size = self.config.vision_config.pixel_shuffle_scale_factor |
| 980 | sizes = spatial_grids.prod(-1) // (merge_size * merge_size) |
| 981 | return tuple(vision_embeddings.split(sizes.tolist())) |
| 982 | |
| 983 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None: |
| 984 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 985 | if image_input is None: |
| 986 | return () |
| 987 | return self._process_image_input(image_input) |
| 988 | |
| 989 | def forward( |
| 990 | self, |
| 991 | input_ids: torch.Tensor | None, |
| 992 | positions: torch.Tensor, |
| 993 | intermediate_tensors: IntermediateTensors | None = None, |
| 994 | inputs_embeds: torch.Tensor | None = None, |
| 995 | **kwargs: object, |
| 996 | ) -> torch.Tensor | IntermediateTensors: |
| 997 | return self.language_model( |
| 998 | input_ids=input_ids, |
| 999 | positions=positions, |
| 1000 | intermediate_tensors=intermediate_tensors, |
| 1001 | inputs_embeds=inputs_embeds, |
| 1002 | **kwargs, |
| 1003 | ) |
| 1004 | |
| 1005 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None: |
| 1006 | return self.language_model.compute_logits(hidden_states) |
| 1007 | |
| 1008 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1009 | loader = AutoWeightsLoader(self) |
| 1010 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1011 | |
| 1012 | def get_mm_mapping(self) -> MultiModelKeys: |
| 1013 | """Get the module prefix in multimodal models.""" |
| 1014 | return MultiModelKeys.from_string_field( |
| 1015 | language_model="language_model", |
| 1016 | connector="vision_embedding.linear_fc2", # The final linear layer |
| 1017 | tower_model="vision_embedding", |
| 1018 | ) |